Computeraided Diagnosisdetection

Computer-aided diagnosis/detection (CAD) can be defined as a diagnosis made by radiologists using the output of a computerized image analysis as an aid. The computer system acts as a "second reader'' by screening the studies and pointing out abnormalities. The final judgment is made by a radiologist. CAD can help improve sensitivity by detecting lesions that might be missed by radiologists.

Figure 13 is the block diagram of a CAD system. A typical CAD system has two phases: training phase and diagnosis phase. In the training phase, a set of training data is first collected. Then radiologist's diagnostic knowledge is applied to analyze the training data and to derive diagnosis rules. In the diagnosis phase, clinical images are taken as input. An image segmentation step is then performed to extract structures of interest. Shape and density features of the structures are computed. A set of features is then selected and fed into a classifier. By using the diagnosis rules obtained in the training phase, the classifier distinguishes actual lesions from false lesions, and hopefully, malignant lesions from benign lesions. The preliminary diagnosis is then reported prior to a final diagnosis.

Features for classification include those traditionally used by radiologists and higher-order features, which may not be very intuitive. Potential features include shape features such as circularity, spherelarity, compactness, irregularity, and elongation, or density features such as contrast, roughness, and texture attributes. Different diagnostic tasks require different sets of useful features. Feature selection techniques, such as forward stepwise methods and genetic algorithms, are applied in the training phase to choose useful feature sets (49). Several classifiers have been proposed for different applications, including linear discriminant analysis, Bayesian methods, artificial neural network, and support vector machine (50).

The quality of a CAD system can be characterized by the sensitivity and specificity of the diagnosis. Sensitivity refers to the fraction of diseased cases

Training phase 1 Diagnosis phase

Figure 13 Block diagram of computer-aided diagnosis.

Training phase 1 Diagnosis phase

Figure 13 Block diagram of computer-aided diagnosis.

correctly identified as positive in the system (true positive fraction, TPF). Specificity refers to the fraction of disease-free cases correctly identified as negative. "Receiver operating characteristic'' (ROC) curves are used to describe the relationship between sensitivity and specificity. The ROC curves show the true-positive fraction (TPF = sensitivity) versus the false-positive fraction (FPF = 1-specificity). In addition to ROC curves, Free-response ROC (FROC) curves (TPF versus false positive per images) were proposed to more accurately represent the number of false positive detections (51). The area under the ROC and FROC curve is a measure of the quality of a CAD system. There is a trade-off between specificity and sensitivity. A successful CAD system should detect as many true lesions as possible meanwhile rejecting as many false positives as possible.

The applications of CAD include the detection of breast microcalcification and mass, pulmonary nodule, and colon polyps. CAD systems are especially useful for the screening of diseases, where a large volume of low incidence examinations need to be screened rapidly. Several companies are currently developing commercial CAD software in mammography, chest radiography and chest CT. R2 Technology Inc's ImageChecker CAD system has received FDA approval for clinical use.

CAD involves all other aspects of image processing techniques. For instance, image segmentation and registration are necessary for feature extraction, and image visualization and measurement are essential for clinical presentation.

DISCUSSION AND CONCLUSION

Medical imaging has become an essential component in many fields of biomedical research and clinical practice. For example, radiologists identify and quantify tumors from MRI and CT images, and neuroscientists detect regional metabolic brain activity from PET and MRI images. Analysis of medical images requires sophisticated image processing techniques. In this chapter, we described several medical image processing fields and emphasized their relevance to tumor imaging researches. The purpose of image segmentation is to localize the tumor regions; image measurement is to quantify the tumor properties; image visualization is to provide intuitive ways to present the tumor; image registration is to fuse two images so that different tumor properties can be combined in one view; finally, CAD could be used in the clinical diagnosis/detection of tumors.

There are quite a few medical image processing and analysis software packages available, both for clinical practices and research activities. Major medical imaging device companies routinely provide high-level image processing workstations to be sold with their imaging equipment. A number of open source or freeware image processing suites are also available. For example, MIPAV is a free software package developed at NIH (52). It is a Java-based application, which can run on any computer platform. MIPAV incorporates a lot of advanced image processing techniques in its package. Insight Segmentation and Registration Toolkit (ITK) is an open-source software package developed by several groups organized by National Library of Medicine. It includes cutting-edge segmentation and registration algorithms.

Medical image processing is a multidisciplinary application area, which involves radiologists, scientists, and technologists. Clinicians recognize the problems and applications during their daily clinical practice. Scientists find solutions to the problems and customize existing tools or develop new tools for the applications. Technologists use the image processing tools to process the clinical images. Radiologists apply the image processing outcome in the diagnosis.

Medical image processing has evolved into an established discipline. It is a very active and fast-growing field. Image processing techniques have already shown great potential in detecting and analyzing tumors in clinical images and this trend will undoubtedly continue into the future.

0 0

Post a comment